Skip to content
Language models

JANSORI

a 3.22B bilingual base model

2 min readupdated

JANSORI is a pretrained 3.22-billion-parameter base LLM with a dense architecture created by pathcosmos. It is designed for text completion in Korean and English without chat templates or role tokens. Unlike instruct-tuned models, it is a raw base model intended as a starting point for fine-tuning on custom tasks. It suits researchers and developers working with Korean-English text or studying the behavior of compact models prior to SFT.

Key features

  • Bilingual focus: trained on roughly 98.57 billion tokens split nearly evenly between Korean (Wikipedia, FineWeb-2, HPLT 2.0) and English data, including math (OpenWebMath) and code (GitHub Code).
  • Ready-to-use quantizations in the repo: alongside the original BF16 weights, Q8_0 and Q4_K_M GGUF builds with SHA-256 checksums are included in the same commit.
  • Compact context and vocabulary: supports a context length of up to 4,096 tokens and uses a 65,536-token vocabulary.
  • Transparent breakdown of limitations: the author provides detailed perplexity (PPL) benchmarks, generation speeds, and failure examples across logic, arithmetic, and code.

What you can do

  • Use it as a fine-tuning base: adapt the model via LoRA or full SFT for Korean and English text generation or processing.
  • Run locally with Ollama or vLLM: test text completion on a consumer GPU using GGUF builds, or deploy a server in BF16.

Example

ollama run hf.co/pathcosmos/JANSORI:Q8_0 "한국의 수도는"
ollama run hf.co/pathcosmos/JANSORI:Q4_K_M "한국의 수도는"
Read moreDetails
Holdout metrics: Korean PPL — 9.09 (BF16 / Q8_0) and 9.20 (Q4_K_M); English PPL — 10.22 (BF16 / Q8_0) and 10.32 (Q4_K_M).
VRAM usage during generation (measured on an H200): 8.8 GB for BF16, 5.06 GB for Q8_0, and 3.76 GB for Q4_K_M.
Training ran for 47,000 steps (98,566,144,000 tokens).
Limitations: like other raw base models without fine-tuning, it tends to repeat phrases, struggles with basic arithmetic, and drops indentation in code.
Discuss Telegram 0 141

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 6.0 GB · ready files7.5 GB
  • GGUF Q8_0recommendedweights 3.2 GB · ready files4.7 GB
  • FP8 / INT8weights 3.0 GB4.5 GB
  • GGUF Q4_K_Mweights 1.8 GB · ready files3.3 GB
  • 4-bitweights 1.7 GB3.2 GB

Tip: go 8-bit.GGUF Q8_0 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: GGUF Q8_0 — from RTX 2060 · BF16 — from RTX 5060 Ti 8GB

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Language models: with a 8192-token context.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # pathcosmos/JANSORI — A30, GGUF Q8_0
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -fsSL https://ollama.com/install.sh | sh
    ollama run hf.co/pathcosmos/JANSORI:Q8_0
    Add more LoRAs to JANSORI
  3. 3. Use it

    The chat opens right in the terminal. Leave with /bye, start again with the same ollama run command.

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

New posts about versions of JANSORI (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

pathcosmos/JANSORI
Author
pathcosmos
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

1410

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Language models

xevyo-01

xevyo-01